Re: [PATCH 01/20] perf nsinfo: fix refcounting

From: Arnaldo Carvalho de Melo
Date: Thu Jan 06 2022 - 15:27:54 EST


Em Tue, Jan 04, 2022 at 10:19:42PM -0800, Krister Johansen escreveu:
> On Thu, Jul 15, 2021 at 04:20:24PM -0300, Arnaldo Carvalho de Melo wrote:
> > Em Thu, Jul 15, 2021 at 04:02:16PM -0300, Arnaldo Carvalho de Melo escreveu:
> > > Em Thu, Jul 15, 2021 at 06:07:06PM +0200, Riccardo Mancini escreveu:
> > > > ASan reports a memory leak of nsinfo during the execution of the perf
> > > > test "31: Lookup mmap thread".
> > > > The leak is caused by a refcounted variable being replaced without
> > > > dropping the refcount.
> >
> > > > This patch makes sure that the refcnt of nsinfo is decreased whenever
> > > > a refcounted variable is replaced with a new value.
> >
> > > So, there are multiple fixes in just one patch, I'll split it into
> > > three, no need to resend.
> >
> > > I'll try and check if finding Fixes: for the three is easy, that way
> > > stable@xxxxxxxxxxxxxxx will figure out which of the supported releases
> > > need each of them.
> >
> > The second is at the end of this message, and looking at it I think
> > there is a problem with where this nsinfo reference is stored, it should
> > be in the 'struct map', that is per-thread, not in the 'struct dso',
> > that may be shared by multiple 'struct map's in different cgroups, so we
> > can't set dso->nsinfo in map__new(), right? Kristen?
>
> Apologies for the long delayed reply, and for the hassle that this has
> caused.
>
> You're right, Arnaldo, that namespace mapping could differ on a per
> thread basis. If the dso objects are intended to be read by multiple
> threads with different views of the target process's namespaces. With
> containers, though, there might be many different versions of the same
> object with the same filename, and the only ways we can disambiguate are
> through the build-ids or by ensuring that the dso refers to a single
> pid and mount namespace. I may have erred in assuming that all of
> perf's threads would be able to access the target namespace via the same
> identifiers. Is that what's going on here?

yeah, I think using build-ids, when available is the way to go, and
since we can almost take it for granted (haven't checked) that it will
be present, as I think gcc/clang adds them by default, seems the way to
go.

That or somehow have the DSO list to be perf namespace.

- Arnaldo

> -K
>
>
> > commit 280f3933661f9ca9563ed04ce82313fdbcbb8f75
> > Author: Riccardo Mancini <rickyman7@xxxxxxxxx>
> > Date: Thu Jul 15 18:07:06 2021 +0200
> >
> > perf map: Fix dso->nsinfo refcounting
> >
> > ASan reports a memory leak of nsinfo during the execution of
> >
> > # perf test "31: Lookup mmap thread"
> >
> > The leak is caused by a refcounted variable being replaced without
> > dropping the refcount.
> >
> > This patch makes sure that the refcnt of nsinfo is decreased whenever a
> > refcounted variable is replaced with a new value.
> >
> > Signed-off-by: Riccardo Mancini <rickyman7@xxxxxxxxx>
> > Fixes: bf2e710b3cb8445c ("perf maps: Lookup maps in both intitial mountns and inner mountns.")
> > Cc: Ian Rogers <irogers@xxxxxxxxxx>
> > Cc: Jiri Olsa <jolsa@xxxxxxxxxx>
> > Cc: Krister Johansen <kjlx@xxxxxxxxxxxxxxxxxx>
> > Cc: Mark Rutland <mark.rutland@xxxxxxx>
> > Cc: Namhyung Kim <namhyung@xxxxxxxxxx>
> > Cc: Peter Zijlstra <peterz@xxxxxxxxxxxxx>
> > Link: http://lore.kernel.org/lkml/55223bc8821b34ccb01f92ef1401c02b6a32e61f.1626343282.git.rickyman7@xxxxxxxxx
> > Signed-off-by: Arnaldo Carvalho de Melo <acme@xxxxxxxxxx>
> >
> > diff --git a/tools/perf/util/map.c b/tools/perf/util/map.c
> > index 8af693d9678cefe0..72e7f3616157ead4 100644
> > --- a/tools/perf/util/map.c
> > +++ b/tools/perf/util/map.c
> > @@ -192,6 +192,8 @@ struct map *map__new(struct machine *machine, u64 start, u64 len,
> > if (!(prot & PROT_EXEC))
> > dso__set_loaded(dso);
> > }
> > +
> > + nsinfo__put(dso->nsinfo);
> > dso->nsinfo = nsi;
> >
> > if (build_id__is_defined(bid))
> >

--

- Arnaldo